Identifying spatial clusters of flood exposure to support decision making in risk management

Identifying spatial clusters of

fl

ood exposure to support decision making

in risk management

Veronika Röthlisberger

⁎

, Andreas P. Zischg

, Margreth Keiler

University of Bern, Institute of Geography, Hallerstrasse 12, CH-3012 Bern, Switzerland

b_{University of Bern, Oeschger Centre for Climate Change Research, Mobiliar Lab for Natural Risks, Falkenplatz 16, CH-3012 Bern, Switzerland} c

University of Bristol, School of Geographical Sciences, University Road, BS8 1SS Bristol, United Kingdom

H I G H L I G H T S

•Evaluation of methods and parameters pertinent toflood exposure analyses

•Spatial exposure analyses support prior-itization inflood risk management.

•Feasible detection of hotspots at nation-al scnation-ale based on spatination-ally explicit data

• Complementary spatial distribution of exposuredensitiesandratiosin Switzer-land

•Data aggregation scheme (i.e. by munic-ipalities or grids) influences the results.

G R A P H I C A L A B S T R A C T

a b s t r a c t

a r t i c l e i n f o

Article history:

Received 31 October 2016

Received in revised form 8 March 2017 Accepted 22 March 2017

Available online xxxx Editor: R. Ludwig

A sound understanding offlood risk drivers (hazard, exposure and vulnerability) is essential for the effective and efficient implementation of risk-reduction strategies. In this paper, we focus on‘exposure’and study the infl u-ence of different methods and parameters offlood exposure analyses in Switzerland. We consider two types of exposure indicators and two different spatial aggregation schemes: the density of exposed assets (exposed num-bers per km2_{) and the ratios of exposed assets (share of exposed assets compared to total amount of assets in a}

specific region) per municipality and per grid cells of similar size as the municipalities. While identifying high densities of exposed assets highlights priority areas for cost-efficient strategies, high exposure ratios can suggest areas of interest for strategies focused on the most vulnerable regions, i.e. regions with a low capacity to cope with a disaster. In Switzerland, the spatial distribution of high exposure densities and exposure ratios tend to be complementary. With regards to the methods, wefind that the spatial cluster analysis provides more informa-tion for the prioritizainforma-tion offlood protection measures than‘simple’maps of spatially aggregated data represent-ed in quantiles. In addition, our study shows that the data aggregation scheme influences the results. It suggests that the aggregation based on grid cells supports the comparability of different regions better than aggregation based on municipalities and is, thus, more appropriate for nationwide analyses.

© 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http:// creativecommons.org/licenses/by-nc-nd/4.0/). Keywords:

Flood exposure Flood risk management Prioritization strategies Spatial cluster analysis MAUP

Switzerland

⁎ Corresponding author at: University of Bern, Institute of Geography, Hallerstrasse 12, CH-3012 Bern, Switzerland.

E-mail addresses:[email protected](V. Röthlisberger),[email protected](A.P. Zischg),[email protected](M. Keiler).

http://dx.doi.org/10.1016/j.scitotenv.2017.03.216

0048-9697/© 2017 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). Contents lists available atScienceDirect

Science of the Total Environment

1. Introduction

Flood risk has been increasing during the last decades on a global scale (IPCC, 2012); this is exemplified by the occurrence offlood events associated with high losses in Europe (e.g. 2002 Danube, Elbe and Vltava catchments, 2007 United Kingdom, 2014 Southeast Europe, 2016 Northwest Europe). Theflood events prompted political actions with a focus on the generation offlood risk maps and enhanced national risk management strategies, e.g. the European Parliament's Floods Directive (2007)or the respective frameworks in Switzerland (Bründl et al., 2009; PLANAT, 2005). Flood risk analysis combines information about the hazard (i.e. the frequency and magnitude offloods), exposure (i.e. the population and assets located inflood-prone areas) and vulner-ability (i.e. the susceptibility of the exposed elements to the hazard) (Klijn et al., 2015; Merz and Thieken, 2004; Papathoma-Köhle et al., 2011; UNISDR, 2015a). These three main factors of the risk analysis show spatiotemporal patterns (Aubrecht et al., 2013; Black and Burns, 2002; Fuchs et al., 2013; Mazzorana et al., 2012; Winsemius et al., 2016. In particular, several studies and reports identify increasing in ex-posure as the main driver of increasing risk (Hallegatte et al., 2013; IPCC, 2012; de Moel et al., 2011). In the future,flood risk will continue to in-crease because of socio-economic development and climate change (Visser et al., 2014; Winsemius et al., 2015). Consequently, effective and efficient strategies for risk reduction are essential for the future (Jongman et al., 2014; Rojas et al., 2013) and a sound understanding of relevant risk drivers is a prerequisite for the implementation of risk-reduction strategies (IPCC, 2012; UNISDR, 2015b).

In this paper, we focus on exposure and how the associated data analysis can influence decisions in different risk-based strategies. Expo-sure analysis is strongly dependent on availability, resolution and qual-ity of data, namely data on assets (i.e. including affected people, buildings and infrastructure) and on the nature of the hazards (i.e. flood extent and magnitude). Asset data characteristics, in particular the spatial resolution and the aggregation level, also influence the choice of methods for exposure analysis. Examples of exposure analysis approaches include intersectingflood areas with average asset values based on aggregated land-use classification (e.g. Cammerer et al., 2013; Jonkman et al., 2008; Muis et al., 2015) and spatially explicit in-tersections of building polygons (Figueiredo and Martina, 2016; Fuchs et al., 2015). The latter approach generates high quality and spatially ex-plicit information on exposure, thereby reducing uncertainties if up-scaled to a larger spatial entity.Merz (2006)compares different expo-sure analysis approaches andde Moel et al. (2015)provide an overview concerning spatial scales. Additionally, the levels at which exposed as-sets are aggregated are dependent on data privacy restrictions, data availability and study objectives. Aggregation levels can range from mu-nicipality level (Fuchs et al., 2015; Hallegatte et al., 2013; Huttenlau et al., 2010; Staffler et al., 2008) to NUTS levels for European studies (Lugeri et al., 2010; Lung et al., 2013) and aggregation based on coun-tries (and‘food producing units’) for global studies (Jongman et al., 2012; UNISDR, 2015a). However, due to limited data availability, com-prehensive object-based and therefore spatially explicit analyses are generally restricted to local and regional levels (Huttenlau et al., 2010; Zischg et al., 2013). Since additional information has become increasing-ly available (e.g. on building stock, i.e. existing buildings within a de-fined environment) throughout Europe based on new European regulations, more accurate information on exposed elements can be ob-tained (e.g.Figueiredo and Martina, 2016; Fuchs et al., 2015) and will be used as a basis for decision-making in risk management.

In Switzerland, object-specific information about the building stock andflood hazards map are available nationwide. In this study, we inves-tigate and test the application of different aggregation and normaliza-tion methods on these datasets and highlight their impact on resultant differences to build awareness among relevant decision makers. The legislative framework and the limited funds for protection measures oblige authorities to prioritize the most efficient and effective risk

reduction schemes. Thus, decision makers need to know“which region should risk reduction focus on?”or alternatively,“where are theflood exposure hotspots located?_”. To answer these questions, we propose an approach of spatial cluster analysis based on the aggregation of point data with respect to different spatial units. Spatial cluster analyses are well established in many disciplines (crime, health, archeology, with Snow's (1855)publication on the 1854 cholera outbreak in Soho district of London being to our knowledge thefirst work on spatial clusters), but with limited applications in natural risk analysis and management to date. The few studies on natural hazards that apply spatial cluster anal-yses (e.g.Borden and Cutter, 2008; Fuchs et al., 2012; Kazakis et al., 2015; Su et al., 2011; van der Veen and Logtmeijer, 2005) often use ag-gregated data and rarely consider the influence of the shape and size of the data aggregation units. In our study, we investigate if and to what extent the aggregation scheme influences the results. In other words, we examine the relevance of the still unresolved and thus often ignored Modifiable Areal Unit Problem (MAUP) (Openshaw, 1984; cf. also Section 2.4in this paper).

We further consider two types of exposure indicators: the density of exposed assets (exposed number of assets per km2_{) and the share of}

ex-posed assets (share of exex-posed assets compared to the total number of assets in a specific region). Thefirst indicator, the exposure density, sup-ports risk management strategies that follow the concept of utilitarian-ism (Mill, 2007). Utilitarianutilitarian-ism in natural hazard and risk management means to choose the most cost-efficient measures. Numerous factors in-fluence a measure's efficiency, i.e. the ratio of resource input to the risk reduction output. The density of exposed assets is an example of the aforementioned factors. Provided that all factors are the same except the exposure density, the efficiency of a measure is higher in areas with high densities of exposed assets than in areas with low exposure densities. That is, the density of exposed assets is a meaningful criterion for the selection of measures with respect to cost efficiency. The second exposure indicator, the share of exposed assets, informs strategies which comply with Rawls' concept of justice (Rawls, 1971). The applica-tion of this concept in risk management implies the prioritizaapplica-tion of the most vulnerable areas and people (Johnson et al., 2007). The term‘ vul-nerable’in this context does not refer to the individual physical suscep-tibility of assets in a region, but to the missing capacity of a region to cope with a disaster. We assume an inverse relationship between the share of affected assets and a region's coping capacity. Consequently, we propose that the share of exposed assets in a given spatial unit is used as an indicator of the unit's vulnerability.

The proposed approach of spatial cluster analysis is generally appli-cable, i.e. for different regional and nationalflood exposure surveys. In this paper it is applied and illustrated with the case study of Switzerland.

2. Material and methods

For the analysis offlood exposure, we overlay spatially explicit infor-mation about buildings and inhabitants with data describingflood prone areas. Based on different aggregations of the exposed assets we search for statistically significant hotspots offlood exposure. The follow-ing sections outline the methods applied and describe the datasets used in the Switzerland case study.

2.1. Data on buildings and inhabitants

Two datasets are extracted from (1) a topographic landscape model and (2) from point data on residential buildings and combined to obtain a comprehensive and homogenous, country-wide database of buildings polygons and of residents in Switzerland.

The feature group‘buildings’from the‘Topographic Landscape Model’(TLM) (swisstopo, 2016a, 2016b) contains footprints of all build-ings currently in Switzerland. The TLM building data is highly accurate (10−1_{m), however, the spatial subsets of the data are not updated}

simultaneously, but regionally in phases between the years 2009–2015. In addition, the change of the building data from 2D to 3D representa-tion is completed only for approximately half of the area (swisstopo, 2016c). Consequently, there are regional differences, e.g. in the degree of division and overlaps of building polygons and the date of underlying aerial images. We address the shortcoming regarding the spatially dif-ferent stages of 2D to 3D conversion by merging all adjoining or overlap-ping building polygons, so that terraced houses or apartment blocks are homogeneously represented as one single polygon. Furthermore, we correct invalid geometries. After these preprocessing steps, our dataset includes 2,086,411 non-overlapping building polygons with a total areal footprint of 540 km2.

The federal ‘Buildings and Dwellings statistic’ (Gebäude- und Wohnungsstatistik (GWS)) contains point-referenced data of all build-ings (entrances) for residential use in Switzerland (Federal Statistical Of_fice FSO, 2016). The number of inhabitants (main residence) at the end of 2012 is used in our study. The dataset contains 1,670,054 points with slightly over 8 million (8,057,480) inhabitants assigned to them.

We assign the number of inhabitants to the building footprints by intersecting the GWS point data with building footprint polygons and applying a snapping distance of≤2 m. Within this distance, 97.7% of all points in the GWS can be attributed to a neighboring building poly-gon. For the total number of inhabitants per building, we total the num-ber of inhabitants of all points associated to a specific building polygon. 2.2. Flood maps

We combine two different types offlood maps to define the areas potentially prone to inundation. The main source of data is a compila-tion of all available communal flood hazard maps, which are complemented by a nationwidefloodplain model called‘Aquaprotect’.

The communalflood hazard maps are generated at the local munic-ipal or cantonal level with respect to Swiss national guidelines (Borter, 1999; Loat and Petrascheck, 1997) and include information from histor-ical events, 2Dflood simulations and expert knowledge. Within the pe-rimeter of the communalflood hazard maps, five different hazard classes (‘high’,‘medium’,‘low’,‘residual’and‘no or negligible’) are de-fined by a specific combination of intensity and probability of events. The communalflood hazard maps are widely accepted and used, espe-cially in the planning process offlood protection measures at communal and cantonal levels (Bundesamt für Umwelt BAFU, 2016a). In our study, we use the March 2016 versions available from the 26 Swiss cantons (federal states) and consider the hazard classes‘high’,‘medium’and ‘low’asflood prone areas (i.e. we include areas affected by events up to a return period of 300 years).

67% (1,390,382) of the building polygons (70% of footprint areas, 77% of residents) are located within the perimeters of the communal flood hazard maps. For the buildings located partially or completely out-side of these perimeters (i.e. buildings withoutflood assessment at the local level), the Aquaprotect dataset provided by the Federal Office for the Environment (Federal Office for the Environment, 2008) is used complementarily. This dataset defines inundation areas based on a “geomorphologic regression”approach (Feyen et al., 2003); this ap-proach has been applied in Switzerland on a 25 m × 25 m grid. Technical flood control facilities and catchment areas below 10 km2_{are not}

con-sidered by Aquaprotect (Federal Office for the Environment, 2008). Aquaprotect includes four different layers with recurrence periods of 50, 100, 250 and 500 years. We use the layer with the 250 year return period in addition to the communalflood hazard map's areas of events with a return period of up to 300 years.

2.3. Exposure analysis

The two datasets–of building footprints and offlood prone areas– are spatially intersected within a GIS to assess theflood exposure of buildings and residents. A building (and its inhabitants) is classified as

exposed if the (partial or whole) footprint polygon overlaps with a flood prone area according to the communalflood hazard maps (classes ‘high_’,_‘medium_’and or_‘low_’). If the building footprint is located outside of the perimeter (i.e. it is withoutflood assessment at the local level), it is exposed if it overlaps with the considered Aquaprotect layer. Build-ings that are located on the fringes of the perimeter of the communal flood hazard maps (i.e. partially inside and outside the perimeter) are classified as exposed if they overlap with one of the three considered classes of theflood hazard maps and/or with the respective Aquaprotect layer.

2.4. Spatial aggregation of data and density calculation

We aggregate the number of exposed buildings, associated footprint and inhabitants based on the currently defined Swiss municipal dis-tricts. The values are then normalized by the area of the respective poly-gon to determine the resultant densities of exposed assets. The Swiss municipalities dataset (swisstopo, 2016d) consists of 2312 entities, cov-ering the entire national territory of 41,290 km2. 2294 of these entities are actually territories of municipalities, 16 are cantonal territories (co-inciding mainly with lake areas) and two are‘communanzas’or public areas managed communally by farmers. The sizes of the 2312 polygons range from 0.1 km2_{to 439 km}2_{with a mean value of 18 km}2_{. With}

ref-erence to motivation stated in the introduction, we are interested in in-vestigating whether data aggregation by municipal districts influences data analysis results. That is, we want to determine if the‘Modifiable Areal Unit Problem’(MAUP) impacts the analysis offlood exposure of Swiss municipalities. The MAUP includes two aspects, (1) the scaling and (2) the zonation effect (Openshaw, 1984; cf. alsoCharlton, 2008). The scaling effect describes the observation that analytical results change based on the level of data aggregation (e.g. block census vs mu-nicipal districts vs county level). The zonation effect describes inconsis-tencies in results when the number of areal units (and thus their average size) remains constant, while boundary positions are shifted. In our study, we focus on the zonation effect by creating an arbitrary grid of 4.23 km × 4.23 km cells, covering the entire Swiss territory. The comparison of the results based on these grid cells (2533 cells, each 17.8929 km2_{) with the results based on aggregation by}

municipal-ities (2312 polygons with an areal average of 17.8592 km2_{) supports the}

assessment of how relevant the MAUP zonation effect is. To calculate the densities of exposed buildings, footprint areas and inhabitants in each grid cell, we apply the quadratic kernel function described by Silverman (1986)with a window 6.345 km wide.

We apply the same density calculation procedures to the total build-ing stock (exposed and unexposed buildbuild-ings) and divide the densities of exposed buildings by the densities of the total building stock. We obtain relative exposure ratios of building numbers, footprint areas and inhab-itants per municipality and per grid cell, respectively.

To determine the robustness of the results based on grid cells we shift the arbitrarily set grid by half of its cell width in each direction, i.e. by 2115 m in north-south and east-west respectively, and repeat the described density calculations. The resulting densities and their dis-tributions within this second grid are very similar to ones in thefirst grid. Thus, we proceed solely with thefirst grid.

2.5. Detection of spatial clusters

‘Hotspots’and‘spatial clusters’are defined differently and many rel-evant techniques can be applied to detect these spatial patterns (see Getis, 2008for a historical outline, andLegendre and Legendre, 2012 for a topical overview and mathematical details). From the available range of interpretations, Levine's definition of a hotspot as an“extreme form of spatial autocorrelation”(Levine, 2008) and Knox's definition of spatial clusters as“geographically…bounded group[s] of occurrences …of sufficient size and concentration to be unlikely to have occurred by chance”(Knox, 1989) are of particular interest.

We apply the local spatial autocorrelation statistic Gi*(d), by Getis and Ord (Getis and Ord, 1992; Ord and Getis, 1995), to identify hotspots of_flood exposure in Switzerland. In other words, the aim is to detect sta-tistically significant clusters of high values in terms of densities of ex-posed buildings (numbers, footprint areas and inhabitant), as well as in terms of relative exposure ratios. The con_fidence level is set at 95%, and ford, we usefixed distance bands. We set thefirst distance band just above the maximal distance between centroids of neighbors (i.e. at 21 km for municipalities and at 4.25 km for grid cells) to assign to each municipality or grid cell at least one neighbor. We investigate the effect of the size ofdby increasing it for municipalities to 31.5 km and 42 km and for grid cells to 8.5 km, 12.75 km, 17 km and 21.25 km.

The evaluation of the Gi*(d) statistical framework reveals additional issues with both spatial dependence and multiple testing. While the spatial dependence violates the test requirement of independent fea-tures, the multiple testing leads to a high number of false type I errors, i.e. incorrect rejections of the null hypothesis. Therefore, it is necessary to adjust the critical values of the Gi* statistics. We use the method pro-posed byCaldas de Castro and Singer (2006).

3. Results and discussion

The data resulting from the assignment of point data on inhabitants (GWS dataset) to building footprints (TLM dataset) are presented in Section 3.1, together with the national level results from theflood expo-sure analysis. InSection 3.2, we discuss selected outcomes of the spatial aggregation investigation and of the density and ratio calculations. Section 3.3describes the results of the spatial clusters analysis. The rel-evance of the MAUP is addressed inSection 3.4, and inSection 3.5, we discuss the implications of ourfindings for the prioritization offlood risk reduction measures. The annex presents supplementary results to the selected outcomes described inSection 3.2.

3.1. Exposed assets: buildings and residents

A total of 97.7% (1,631,531) of the GWS data points are assigned to a building footprint polygon (i.e. 93.7% are located within a TLM building footprint polygon and 4% within a distance of≤2 m). These assigned GWS data points represent 98.2% (7,909,191) of the Swiss residents. The high percentage of residents assigned to buildings reflects the high spatial accuracy of both the GWS points and the TLM footprints datasets.

Overall, 320,509 buildings in Switzerland are identified as exposed tofloods up to a return period of 250 to 300 years (seeTable 1). This is equivalent to an exposure ratio of 15.4%. This ratio is in agreement with results from previous studies conducted with Switzerland as a study site (Bundesamt für Umwelt BAFU, 2016b; Fuchs et al., 2017) and is in line withfindings about other mountainous regions in Europe (Chen et al., 2016, Fuchs et al., 2015; Url and Sinabell, 2008). The expo-sure ratios of the building footprint areas and of the number of residents are higher (20.4% and 17.1%, respectively) but still comparable to the ratio regarding the numbers of buildings. This indicates that the average footprint area of exposed buildings is greater than the average footprint area of unexposed buildings in Switzerland (343 m2_{for exposed versus}

244 m2_{for unexposed, cf.Table 1). The same comparison is valid for the}

number of residents. However, the relative difference between the aver-age number of inhabitants in exposed buildings (4.2 persons per

building) and in unexposed buildings (3.7 persons per building) is smaller.Fuchs et al. (2015)obtain similarfindings with respect to larger buildings located inside of_flood prone areas and explain that larger buildings require moreflat terrain. Consequently, they are preferably built onfloodplains. A second explanation for the higher exposure ratios associated with larger buildings is related to the applied method. It scores any building overlapping withflood prone areas as“exposed”, re-gardless of the percentage of the building footprint that intersects with theflood area feature. As a result, large building polygons are more like-ly to be scored as“exposed”than small ones.

The last column inTable 1presents the exposed elements normal-ized by the total area of Switzerland (41,290 km2). An average of eight buildings (2,700 m2_{footprint area, 33 inhabitants) are exposed per}

square kilometer. These normalizedfigures are revisited in the follow-ing section, where we present the densities of exposed assets based on spatial aggregations with different spatial units.

3.2. Spatial aggregates, densities and ratios of exposed assets

Fig. 1presents the aggregation of the number of exposed buildings per municipality normalized per square kilometer infive 20%-quantiles. The limit between the middle and the second highest class coincides with the average value of eight exposed buildings per square kilometer considering all of Switzerland (cf.Table 1). Consequently, 60% of the municipalities (the three lower classes) have densities that are less than or equal to the nationwide average. Furthermore, the mean of the density values of all municipalities is approximately 13 exposed buildings per square kilometer, which isN50% higher than the nation-wide density. In other words, the density distribution per municipality is not symmetrical but right-skewed and small municipalities are gener-ally associated with higher densities than large ones. Considering the spatial distribution of the densities, low values are generally concentrat-ed in the south-eastern part of Switzerland. In contrast, many munici-palities with high densities are located in the central and the northern parts of the country, with the exception of two lines of high values in the south (areas A and B inFig. 1).

Fig. 2also presents densities of the number of exposed buildings, but per 4.23 km × 4.23 km grid cells. The thresholds used to define thefive plotted classes are the same as inFig. 1. Overall, the spatial distribution pattern of the densities per grid cell inFig. 2is similar toFig. 1, but the high density values in area B (inFig. 1) are less prominent. Here, the high values of some small municipalities (≪18 km2) are smoothed out when the densities are calculated per grid cell. This smoothing effect of high and low density values affects all regions, where the areas of mu-nicipalities are considerably smaller than the areas of the grid cells. Ad-ditionally, the applied kernel density estimation with a window width of 6.345 km smoothens the resultant values. However, this smoothing effect of the kernel density estimation is cancelled out in regions where the municipal areas are notably larger than the grid cell areas. Consequently, in regions with large municipalities, the variability of the underlying exposure data is more evident in the density maps orga-nized per grid cell (e.g. area C inFig. 2) than per municipality.

In principal, the exposure density in a particular spatial unit is the re-sult of two underlying features, (1) the density of assets (number of buildings, footprints or inhabitants per square kilometer of this spatial unit) and (2) the proportion offlood prone area to the total area of this spatial unit. While the high exposure density in the northern part

Table 1

Flood exposure in Switzerland. Norm. = normalized per km2

, surface of Switzerland: 41,290 km2 .

All Not exposed Exposed

Asset type [N] or [106 m2 ] [N] or [106 m2 ] [N] or [106 m2 ] Ratio [%] Norm. [/106 m2 ] Number of buildings [N] 2,086,411 1,765,902 320,509 15.4 8

Footprint area of buildings [106 m2

] 540 430 110 20.4 0.0027

of Switzerland is attributed to an overall high density of assets in these lowlands and hilly regions, the high exposure densities in the moun-tainous south correspond with the high settlement concentrations lo-cated on the main alpine river floodplains that extend across the entire valley bottom (Zimmermann and Keiler, 2015).

The spatial distributions of the building exposure ratios per munici-pality are shown inFig. 3, classified infive quantiles of 20%. The 15.4% nationwide exposure ratio (seeTable 1), as well as the mean of the ex-posure ratios of all municipalities (13.9%) belong to the second highest class, indicating a right-skewed distribution. However, unlike the densi-tyvalues per municipality, the exposureratiosper municipality do not show a correlation to the areal size of municipality. Considering the spa-tial distribution of the exposure ratios per municipality, we identify a narrow area of high ratios from the southwest to the northeast of Swit-zerland, which coincide with the northern and pre-alpine regions of the Swiss Alps. Otherwise, low to medium ratios dominate, with the excep-tion of some isolated high values.

Finally,Fig. 4presents the exposure ratios per grid cell with the same thresholds of thefive classes as inFig. 3. Compared withFig. 3 (munic-ipalities) an equivalent spatial distribution of the exposure ratios ap-pears in Fig. 4(grids). As in the case of the exposure ratios per municipality, the exposure ratios per grid cell are generally higher in a broad strip that stretches from the southwest to the northeast. In

addition, we identify an accumulation of high values in the west (area D inFig. 4), which does not appear inFig. 3(ratios per municipalities) due to the presence of some polygons without data. These polygons rep-resent parts of lakes that are not assigned to municipalities and that do not contain buildings; thus, no exposure ratios are calculated. Finally, in the northern part of Switzerland, the map of exposure ratios per grid cells (Fig. 4) is easier to interpret than the map per municipalities (Fig. 3). This is mainly due to the smoothing effects caused by larger grids cells than municipalities and by the kernel density estimation procedure.

The prominent concentration of high exposure ratios in the narrow region that spans from the southwestern to the northeastern parts coin-cides with the northern and central parts of the Swiss Alps. This is a re-gion that includes both concentrated settlements and wide areas characterized by relatively sparse populations. The main settlements here are concentrated in the few relativelyflat areas, which are mainly floodplains, alluvial fans and debris cones.

When comparing exposuredensitieswith exposureratios(both sets of values organized per municipality or per grid cell), the extreme values (in the highest and the lowest classes) are found to be largely complementary. There are only a few major areas that are consistent.

Fig. 1.Densities offlood exposed buildings per Swiss municipality, numbers per km2 . Data classes are (rounded) 20% quantiles. Fine lines represent municipal boundaries and thick lines represent cantonal boundaries. Areas A and B highlight two lines of high values in the south of Switzerland.

Fig. 2.Densities offlood exposed buildings in Switzerland per 4.23 × 4.23 km2_{grid cell,} numbers per km2_{. Thresholds of classes are identical to Fig. 1 (densities per} municipalities). Fine lines represent grid cells and thick lines represent cantonal boundaries. Area C highlights an example, where the variability of exposure is more evident in density maps per grid cell than per municipality.

Fig. 3.Ratios offlood exposed buildings per Swiss municipality [numbers of exposed buildings/numbers of all buildings]. Data classes are (rounded) 20% quantiles. Fine lines represent municipal boundaries and thick lines represent cantonal boundaries. Area E highlights an example, where low values are observed in all four maps with 20% quantile representation (Figs. 1 to 4), and area F is characterized by high values in all four maps.

Fig. 4.Ratios offlood exposed buildings in Switzerland per 4.23 × 4.23 km2_{grid cell} [numbers of exposed buildings/numbers of all buildings]. Thresholds of classes are identical toFig. 3(ratios per municipalities). Fine lines represent grid cells and thick lines represent lines cantonal boundaries. Area D highlights an accumulation of high values, which does not appear inFig. 3(ratios per municipalities).

Two examples are highlighted inFig. 3: Area E is a region where low values are observed in all four maps (Figs. 1 to 4, see for a better over-view Fig. A1 in Appendix) and area F is characterized by high values in all four maps. Overall, areas with high values in all four maps corre-spond with the relatively wide valleys of the main rivers connecting the high alpine regions with the foothills of the Alps, in particular, the Rhine river valley upstream of Lake Constanze (area F inFig. 3) and the Rhone valley upstream of Lake Geneva (area A inFig. 1). Both are considered to be densely populated areas surrounded by rather sparsely populated mountain areas. The large region of low values in the south-east observed in all four maps (Fig. A1 in Appendix) is almost congruent with the territory of Canton Graubünden. Authorities of Canton Graubünden pioneered the consideration of natural hazard aspects in the spatial planning processes in Switzerland. Legally binding hazard maps were introduced as early as 1963.

The aforementioned statements regarding the numbers of buildings also generally apply to the results describing the areal building foot-prints, and to a lesser degree, to the inhabitants. To provide an overview of the spatial distributions, the appendix presents the aggregations of the numbers of buildings (Fig. A1), of the areal building footprints (Fig. A2), and of the number of inhabitants (Fig. A3).

3.3. Hotspots offlood exposure

The results of the hotspot analyses are summarized inFigs. 5 and 6. The dark colored areas in all of the maps represent clusters of high values, based on the Gi*(d) statistic (Getis and Ord, 1992,Ord and Getis, 1995) and on a 95% confidence level.Fig. 5provides an overview of 30 maps based on thefive distance bandsd(applied on data aggre-gated on grid cells, see columns inFig. 5) and on the six analyzed expo-sure indicators (rows inFig. 5) aggregated per grid cells. The results show (Fig. 5) that an increasing distance band value results in larger and more generalized hotspots, but does hardly change the position of these areas. The 18 maps inFig. 6provide a comparable overview to the one presented inFig. 5and are based on the same six indicators (rows inFig. 6), but aggregated per municipalities and based on three distance bands d (columns inFig. 6). In contrast to the grid cell aggrega-tion, the results of the municipality-based aggregation highlight that an increasing distance band not only changes the size and shape of hotspot areas but, also their position under some circumstances (e.g. in the cen-tral northern part in thefirst row inFig. 6). Thus, due to the spatially more stable results obtained, the grid cell approach is considered to be more appropriate for hotspot analyses. In addition, we consider the

Fig. 5.Hotspots offlood exposure in Switzerland, based on data aggregated on 4.23 × 4.23 km2_{grid cells. The dark colored areas show statistically significant clusters of high values (see} legend) based on the local spatial autocorrelation statistic Gi*(d) by Getis and Ord (references see text), forfive differentfixed distance bands (columns) and six different types of indicators (rows). Confidence level at 95%, correction of false discovery rate applied. Details on the applied method described inSection 2.5. The two maps, which are replicated inFig.7, are highlighted by a bold frame. Gray lines represent cantonal boundaries.

analyses based on grid cells as‘spatially more consistent’in the sense that the sizes of the cells are all identical. Consequently, the size factor does not influence the result the way that variable sizes of municipali-ties do. For instance, the high exposure ratio (numbers of buildings) in the most eastern municipality of Switzerland is not identified as a part of a cluster (see_first row with brown colored maps inFig. 6), simply be-cause this municipality is comparably large and has no neighboring mu-nicipalities with high exposure ratios (seeFig. 3). If the underlying data of the same municipality are aggregated on smaller areal units (e.g. grid cells of approximately 18 km2_{), the same hotspot analysis shows}

signif-icant clusters of high exposure ratios (seefirst row with brown colored maps inFig. 5). While the reason that data aggregation on grid cell is to prefer over the aggregation based on administrative boundaries is gen-erally valid, the optimal size of the distance banddis dependent on the

purpose of the spatial cluster analysis. It represents a compromise be-tween producing continuous areas (by increasingd) and maintaining spatial differentiations (by decreasingd). Based on the evaluation of the parameters, we consider hotspot analyses based on grid cells and with a distance band of 17 km (second last column inFig. 5). The select-ed parameters are optimal for providing a nationwide overview. Further analysis follows this optimal approach.

Fig. 7presents hotspots of the number of exposed buildings, based on data aggregated on 4.23 km × 4.23 km grid cells and with a distance band of 17 km. It can be observed that the hotspots based on density values are highly complementary to the hotspots based on exposure ra-tios. While the hotspots based on density values are mainly located in the northern part of Switzerland (with some additional spots in the southwest), the majority of the ones based on exposure ratios are

Fig. 6.Hotspots offlood exposure in Switzerland, based on data aggregated on municipalities. The dark colored areas show statistically significant clusters of high values (see legend) according to local spatial autocorrelation statistic Gi*(d) by Getis and Ord (references see text), for three differentfixed distance bands (columns) and six different types of indicators (rows) Confidence level at 95%, correction of false discovery rate applied. Details on the applied method described inSection 2.5. Gray lines represent cantonal boundaries. (For interpretation of the references to color in thisfigure legend, the reader is referred to the web version of this article.)

located within a broad strip from the southwestern to the northeastern parts of the country. Additionally, there are two small and isolated hotspots of exposure ratios, one located in the west and another in the east. Dark gray colored areas inFig. 7are areas where hotspots based on exposure densities and hotspots based on exposure ratios overlap. These dark gray areas are comparatively small and limited to the inter-mediate zones between the high alpine regions and the foothills of the Alps.

The spatial distributions of hotspots based on building footprint or on inhabitants respectively are similar to the distribution of hotspots based on the number of buildings (seeFigs. 5 and 6).

3.4. Relevance of the MAUP

Our results show that the MAUP is relevant. The aggregation scheme influences both groups of results, namely, the spatial aggregates (see Section 3.2) and the hotspots (seeSection 3.3) offlood exposed assets. In particular, the aggregation scheme does not influence thegeneral spatial distributions offlood exposure densities or ratios with respect to spatial aggregates. They remain the same whether the data is aggregat-ed basaggregat-ed on municipality districts or regular grids. However, upon fur-ther detailed inspection, they differ due to the effects of smoothing and coarsening (seeSection 3.2). These effects illustrate the scaling as-pect of the MAUP. In relation to MAUP, there are two reasons why ag-gregation based on grid cells is considered to be more appropriate for nationwide analyses. Firstly, aggregation based on (regular) grid cells supports the comparability of different regions better than the aggrega-tion based on (irregularly shaped and sized) municipalities. Secondly, in regions with (very) small municipal districts (e.g. in the northwestern parts or in the very south of Switzerland), the visualization of aggregat-ed values basaggregat-ed on grid cells provides a more usable overview of the sit-uation at the presented scale.

The comparison of the respective hotspots based on the two differ-ent aggregation schemes shows that they differ not so much with regards to exposure ratios (brown colored maps inFigs. 5 and 6), but re-markably when looking at exposure densities (blue colored maps in Figs. 5 and 6). The two different aggregation schemes using similar numbers and average sizes of spatial units, result in different areas of

flood exposure hotspot densities in Switzerland. That is, the results of these hotspot analyses are inconsistent when applying the two aggrega-tion schemes, which vary in border locaaggrega-tions. In other words, it matters where the borders of the spatial units are drawn. Consequently, the MAUP, especially the zonation aspect, should not be neglected. 3.5. Prioritization in risk reduction and risk management strategies

The dark colored areas in all presented maps (i.e. in the maps with 20% quantiles and in the hotspots maps) address the initial question pertaining to whereflood protection measures should be prioritized. The dark blue areas (of exposure densities) indicate priorities for strat-egies following the concept of utilitarianism (Mill, 2007), whereas the dark brown ones (of high exposure ratios) identify priority areas for strategies which comply with Rawls' concept of justice (Rawls, 1971). In risk management, utilitarianism requests prioritization of the most cost efficient measures. High cost efficiency in turn is linked to high den-sity of exposed assets, which is represented by the dark blue areas in this study. In contrast to utilitarianism, Rawls' concept on justice is fo-cused on the most vulnerable areas and people. Vulnerable means to have low capacity to cope with a disaster. In this study, we use the share of exposed assets as an indicator for vulnerability, represented by brown colors in our maps. The maps with 20% quantile representa-tion (of exposure densities and ratios, Figs. A1 to A3) are limited to the single spatial units of data aggregation, which are municipal districts or grid cells in our study. However,_flood protection measures are often more effective when covering more than a single municipal district or a grid cell of 4.23 by 4.23 km (Thaler et al., 2016). Consequently, for strat-egy prioritization, there is less interest in the high values ofsinglespatial units and greater interest inclustersof high values, i.e. in hotspots. As hotspots represent statistically significant spatial clusters (in our case at 95% con_fidence level) of high values, they support an evidence-based prioritization of regions for protection measures. Hence, the dark colored areas inFig. 7suggest regions of national priority for pro-tection measures.

Switzerland'sflood risk management strategy is still mainly driven by the reaction after large damagingflood events (Suter et al., 2016). The same reaction was observed byThieken et al. (2016)after the 2013flood in Germany and also in other countries. However, the num-ber of preventive risk reduction projects in Switzerland are increasing with an emphasis on areas, which our study identifies as hotspots based on exposure densities. This emphasis re_flects a commonly recom-mended strategic focus on cost efficiency (Meyer et al., 2013; Mori and Perrings, 2012). It is implemented into the decision making process on the allocation of federal subsidies by a respective tool to provide evi-dence of cost efficient measures (Bründl et al., 2009). Only measures with evident cost efficiency are supported by national funds. Currently, there is increasing awareness about certain regions indicated in our study as the most vulnerable territories within Switzerland. However, the focus in these areas is on their capacity to cope withfloods (and other natural hazards), rather than on the reduction of their exposure ratios. Thus, it would be beneficial to determine whether current risk management strategies e.g. intervention or risk transfer by extensive in-surance systems (Gretener, 2011; von Ungern-Sternberg, 2004) in these vulnerable areas are suf_ficient in the case of the occurrence of an event (as discussed within the Austrian context byHolub and Fuchs, 2009).

Nevertheless, within Switzerland's strategic focus on cost efficiency, the most vulnerable regions identified in our study are not neglected. For instance, the federal authorities support precisely two current tech-nicalflood protection projects, with a dedicated contact person. These projects are the‘Alpenrhein Expansion Project’conducted on the river Rhine upstream of Lake Constance and the‘Third Rhone Correction’ up-stream of Lake Geneva (Federal Office for the Environment, 2016). These two projects are not only operating within identified hotspots based on density values, but also within hotspots based on exposure ra-tios. As a result, the aforementioned projects lay within the dark gray

Fig. 7.(Selection and intersection of the two highlighted maps inFig. 5): Hotspots of numbers offlood exposed buildings in Switzerland, based on data aggregated on 4.23 × 4.23 km2

grid cells. The dark colored areas show statistically significant clusters of high values (see legend) according to local spatial autocorrelation statistic Gi*(d) by Getis and Ord (references see text). Confidence level at 95%,fixed distance band at 17 km, correction of false discovery rate applied. Details on the applied method described inSection 2.5. Red ellipses highlight the location of two majorflood protection project in Switzerland. (For interpretation of the references to color in this

colored areas inFig. 7(see red ellipses inFig. 7) that indicate areas where the hotspots based on density values and on ratios overlap. This means that the prioritization of these two projects at the national level is in line with our insights. Ourfindings suggest that the prioritiza-tion of these regions of overlapping hotspots is required to support the realization of both types of strategies focused on cost ef_ficiency and on the most vulnerable regions.

4. Conclusions and outlook

The preceding sections illustrate the utility and pitfalls of spatial sta-tistics applied onflood exposure data in Switzerland. We show that the detection of hotspots, i.e. of statistically significant clusters of high values, is feasible at the national scale and based on the use of spatially explicit data. We_find that spatial cluster analyses support the genera-tion of more informative databases, which can be used to prioritize flood protection measures, especially compared with the limited infor-mation from‘simple’maps of spatially aggregated data represented in quantiles. However, the analysis results into more than one single an-swer to the question‘where are the hotspots offlood exposure?’, at least in the case of Switzerland. Thus, the proposed analysis provides a broad basis for decisions on different types of prioritization strategies inflood risk management.

First of all, the answer depends on the type of the indicator. The re-sults of our case study suggest largely complementary hotspots based on exposure densities and exposure ratios. This means that priority areas for protection measures following cost efficient strategies (utili-tarianism) and for measures focusing on the most vulnerable regions (Rawls' concept on justice) hardly overlap in Switzerland. Identifying these differences on a national level could already be an important step towards evaluating prioritization strategies. The prioritization of cost efficient measures is a well-established strategy with respective tools and criteria supporting the decision process in Switzerland and elsewhere. The density of exposed assets is a key determinant of cost ef-ficiency. In contrast, the focus on vulnerability is less common and the development of the respective concepts and tools is still at a very early stage. The term vulnerability is already subject to ongoing aca-demic discussions (cf.Birkmann et al., 2013), even more diverse are the existing concepts of vulnerability assessments upon whichflood management decisions are based. The exposure ratio, used in this study, is one of all conceivable criteria for vulnerability assessments.

Secondly, there may be differences in the answer when considering different kinds of assets. However, only minor differences between the results regarding the number or footprint or inhabitants of exposed buildings are identified in Switzerland. More importantly, the way data aggregation is conducted influences the results. That is, the MAUP is relevant and must not be neglected in any spatial cluster anal-ysis based on aggregatedflood exposure data. By presenting hotspots based on different distance bands, we further exemplify the influence of parameter settings on the results of a hotspot test statistic. Not only the parameter setting influence the identification of spatial clusters, but already the type of the test statistic that is applied does. In future studies, it might be interesting to apply other spatially explicit local sta-tistics and to compare them with the presented approaches, e.g. proce-dures presented byAldstadt and Getis (2006),Anselin (1995)orTango and Takahashi (2005). Data aggregation based on small catchment areas (instead of grid cells or municipalities), combined with the use of connectivity indices (instead of Euclidean distance) as the neighbor-hood criteria, would be another promising approach for future improvements.

Regardless of the data and the methods used, it is essential to select them based on the questions to be answered forflood risk management. Furthermore, we emphasize the utility of publishing hotspots offlood exposure in combination with notes on their dependency on the param-eters of the applied method. This way,flood exposure hotspot analyses

provide added value to evidence-based decisions making pertaining to the prioritization offlood risk reduction measures.

Supplementary data to this article can be found online athttp://dx. doi.org/10.1016/j.scitotenv.2017.03.216.

Acknowledgments

Funding from the Swiss Mobiliar (_‘die Mobiliar AG_’) supported the completion of this research. We thank the natural hazards group at the Swiss Mobiliar for acquiring and compiling the communalflood hazard maps; Candace Chow for editorial assistance; Sven Fuchs for proofreading and Markus Mosimann for his support with the design of thefigures. We like to thank the anonymous reviewer and the editor for their inputs which substantially improved this manuscript.

References

Aldstadt, Jared, Getis, Arthur, 2006. Using AMOEBA to create a spatial weights matrix and identify spatial clusters. Geogr. Anal. 38 (4):327–343.http://dx.doi.org/10.1111/j. 1538-4632.2006.00689.x.

Anselin, Luc, 1995.Local indicators of spatial association - LISA. Geogr. Anal. Vol 27 (2), 93–115.

Aubrecht, Christoph, Fuchs, Sven, Neuhold, Clemens, 2013. Spatio-temporal aspects and dimensions in integrated disaster risk management. Nat. Hazards 68 (3): 1205–1216.http://dx.doi.org/10.1007/s11069-013-0619-9.

Birkmann, Jörn, Cardona, Omar D., Carreño, Martha L., Barbat, Alex H., Pelling, Mark, Schneiderbauer, Stefan, et al., 2013. Framing vulnerability, risk and societal re-sponses. The MOVE framework. Nat. Hazards 67 (2):193–211.http://dx.doi.org/10. 1007/s11069-013-0558-5.

Black, Andrew R., Burns, John C., 2002. Re-assessing theflood risk in Scotland. Sci. Total Environ. 294 (1–3):169–184.http://dx.doi.org/10.1016/S0048-9697(02)00062-1. Borden, Kevin A., Cutter, Susan L., 2008. Spatial patterns of natural hazards mortality in the

United States. Int. J. Health Geogr. 7:64.http://dx.doi.org/10.1186/1476-072X-7-64. Borter, Patricio, 1999. Risikoanalyse bei gravitativen Naturgefahren. Methode.Edited by

Bundesamt für Umwelt, Wald und Landschaft (BUWAL) (Umwelt-Materialien, Nr. 107/I). Available online at.http://www.bafu.admin.ch/publikationen/publikation/ 00131/index.html?lang=de(checked on 7/21/2016).

Bründl, Michael, Romang, Hans E., Bischof, Nicole, Rheinberger, Christoph M., 2009.The risk concept and its application in natural hazard risk management in Switzerland. Nat. Hazards Earth Syst. Sci. 9, 801–813.

Bundesamt für Umwelt BAFU, 2016a.Von der Risikoanalyse zur Massnahmenplanung. Arbeitsgrundlage für Hochwasserschutzprojekte Bern (Umweltwissen, 1606). Bundesamt für Umwelt BAFU, 2016b.Umgang mit Naturgefahren in der Schweiz - Bericht

des Bundesrates.

Caldas de Castro, Marcia, Singer, Burton H., 2006.Controlling the false discovery rate: a new application to account for multiple and dependent tests in local statistics of spa-tial association. Geogr. Anal. 38, 180–208.

Cammerer, Holger, Thieken, Annegret H., Verburg, Peter H., 2013. Spatio-temporal dy-namics in theflood exposure due to land use changes in the Alpine Lech Valley in Tyrol (Austria). Nat. Hazards 68 (3):1243–1270. http://dx.doi.org/10.1007/s11069-012-0280-8.

Charlton, Martin, 2008.Modifiable areal unit problem (MAUP). In: Kemp, Karen K. (Ed.), Encyclopedia of Geographic Information Science. SAGE Publications, Inc., pp. 289–290.

Chen, Lixia, van Westen, Cees J., Hussin, Haydar, Ciurean, Roxana L., Turkington, Thea, Chavarro-Rincon, Diana, Shrestha, Dhruba P., 2016. Integrating expert opinion with modelling for quantitative multi-hazard risk assessment in the Eastern Italian Alps. Geomorphology 273:150–167.http://dx.doi.org/10.1016/j.geomorph.2016.07.041. de Moel, Hans, Aerts, Jeroen C.J.H., Koomen, Eric, 2011. Development offlood exposure in

the Netherlands during the 20th and 21st century. Glob. Environ. Chang. 21 (2): 620–627.http://dx.doi.org/10.1016/j.gloenvcha.2010.12.005.

de Moel, Hans, Jongman, Brenden, Kreibich, Heidi, Merz, Bruno, Penning-Rowsell, Edmund, Ward, Philip J., 2015. Flood risk assessments at different spatial scales. Mitig. Adapt. Strateg. Glob. Chang. 20 (6):865–890.http://dx.doi.org/10.1007/ s11027-015-9654-z.

European Parliament, 2007.Directive 2007/60/EC of the European Parliament and of the Council of 23 October 2007 on the assessment and management offlood risk. 2007/ 69/EC. Off. J. Eur. Union 288, 27–34.

Federal Office for the Environment FOEN, 2008. Terms of Use Aquaprotect.Available on-line at.http://www.bafu.admin.ch/naturgefahren/14186/14801/16598/16600/index. html?lang=en(checked on 9/26/2016).

Federal Office for the Environment FOEN, 2016. Flood Protection Section. Responsibilities of Staff.Available online at.http://www.bafu.admin.ch/org/organisation/00180/ 00188/index.html?lang=en(checked on 10/2/2016).

Federal Statistical Office FSO, 2016. Buildings and Dwellings Statistic (Since 2009). Factsheet - Surveys, Sources. Neuchâtel.Available online at.https://www.bfs.admin. ch/bfs/en/home/statistics/construction-housing/surveys/gws2009.html(checked on 1/9/2017).

Feyen, Hans, Mehlhorn, Jens, Menzinger, Ivo, 2003.Invention: Method for Evaluating Flood Plain Risks (Patent Application in the United States Patent and Trademark Office).

Figueiredo, Rui, Martina, Mario L.V., 2016. Using open building data in the development of exposure data sets for catastrophe risk modelling. Nat. Hazards Earth Syst. Sci. 16 (2): 417–429.http://dx.doi.org/10.5194/nhess-16-417-2016.

Fuchs, Sven, Keiler, Margreth, Sokratov, Sergey, Shnyparkov, Alexander, 2013. Spatiotem-poral dynamics. The need for an innovative approach in mountain hazard risk man-agement. Nat. Hazards 68 (3):1217–1241. http://dx.doi.org/10.1007/s11069-012-0508-7.

Fuchs, Sven, Keiler, Margreth, Zischg, Andreas P., 2015. A spatiotemporal multi-hazard ex-posure assessment based on property data. Nat. Hazards Earth Syst. Sci. 15 (9): 2127–2142.http://dx.doi.org/10.5194/nhess-15-2127-2015.

Fuchs, Sven, Ornetsmüller, Christine, Totschnig, Reinhold, 2012. Spatial scan statistics in vulnerability assessment. An application to mountain hazards. Nat. Hazards 64 (3): 2129–2151.http://dx.doi.org/10.1007/s11069-011-0081-5.

Fuchs, Sven, Röthlisberger, Veronika, Thaler, Thomas, Zischg, Andreas P., Keiler, Margreth, 2017.Natural hazard management from a co-evolutionary perspective: exposure and policy response in the European Alps. Ann. Am. Assoc. Geogr. 107 (2).

Getis, Arthur, Ord, John K., 1992.The analysis of spatial association by use of distance sta-tistics. Geogr. Anal. Vol. 24 (3), 189–206.

Getis, Arthur, 2008.A history of the concept of spatial autocorrelation: a geographer's per-spective. Geogr. Anal. 40, 297–309.

Gretener, Max, 2011. Die Versicherung von Elementarschäden durch die privaten Sachversicherer in der Schweiz. Schweizerischer Versicherungsverband, Zürich Available online at. http://www.svv.ch/de/publikationen/die-versicherung-von-elementarschaeden-durch-die-privaten-sachversicherer-der-schweiz(checked on 8/22/2014).

Hallegatte, Stephane, Green, Colin, Nicholls, Robert J., Corfee-Morlot, Jan, 2013. Future

flood losses in major coastal cities. Nat. Clim. Chang. 3 (9):802–806.http://dx.doi. org/10.1038/nclimate1979.

Holub, Markus, Fuchs, Sven, 2009. Mitigating mountain hazards in Austria–legislation, risk transfer, and awareness building. Nat. Hazards Earth Syst. Sci. 9 (2):523–537. http://dx.doi.org/10.5194/nhess-9-523-2009.

Huttenlau, Matthias, Stötter, Johann, Stiefelmeyer, Heinz, 2010. Risk-based damage po-tential and loss estimation of extremeflooding scenarios in the Austrian Federal Province of Tyrol. Nat. Hazards Earth Syst. Sci. 10 (12):2451–2473.http://dx.doi. org/10.5194/nhess-10-2451-2010.

IPCC, 2012.Managing the Risks of Extreme Events and Disasters to Advance Climate Change Adaptation. A Special Report of Working Groups I and II of the Intergovern-mental Panel on Climate Change. With assistance of Field, C.B., V. Barros, T.F. Stocker, D. Qin, D.J. Dokken, K.L. Ebi, M.D. Mastrandrea, K.J. Mach, G.-K. Plattner, S.K. Allen, M. Tignor, P.M. Midgley. IPCC. Cambridge University Press, Cambridge, UK and New York, NY, USA.

Johnson, Clare, Penning-Rowsell, Edmund, Parker, Denis, 2007. Natural and imposed in-justices: the challenges in implementing 'fair'flood risk management policy in En-gland. Geogr. J. 173:374–390http://www.jstor.org/stable/30130632.

Jongman, Brenden, Hochrainer-Stigler, Stefan, Feyen, Luc, Aerts, Jeroen C.J.H., Mechler, Reinhard, Botzen, Wouter W.J., et al., 2014. Increasing stress on disas-ter-riskfinance due to largefloods. Nat. Clim. Chang. 4 (4):264–268 Available on-line at.http://www.nature.com/nclimate/journal/v4/n4/full/nclimate2124.html (checked on 7/2/2014).

Jongman, Brenden, Ward, Philip J., Aerts, Jeroen C.J.H., 2012. Global exposure to river and coastalflooding. Long term trends and changes. Glob. Environ. Chang. 22 (4): 823–835.http://dx.doi.org/10.1016/j.gloenvcha.2012.07.004.

Jonkman, Bas S.N., Bočkarjova, Marija, Kok, Matthijs, Bernardini, Patrizia, 2008. Integrated hydrodynamic and economic modelling offlood damage in the Netherlands. Ecol. Econ. 66 (1):77–90.http://dx.doi.org/10.1016/j.ecolecon.2007.12.022.

Kazakis, Nerantzis, Kougias, Ioannis, Patsialis, Thomas, 2015. Assessment offlood hazard areas at a regional scale using an index-based approach and Analytical Hierarchy Pro-cess: application in Rhodope-Evros region, Greece. Sci. Total Environ. 538:555–563. http://dx.doi.org/10.1016/j.scitotenv.2015.08.055.

Klijn, Frans, Kreibich, Heidi, de Moel, Hans, Penning-Rowsell, Edmund, 2015. Adaptive

flood risk management planning based on a comprehensiveflood risk conceptualisa-tion. Mitig. Adapt. Strateg. Glob. Chang. 20 (6):845–864.http://dx.doi.org/10.1007/ s11027-015-9638-z.

Knox, E. George, 1989.Detection of clusters. In: Elliott, Paul (Ed.), Methodology of En-quiries Into Disease Clustering, pp. 17–22 Workshop. Papers and discussions. (London).

Legendre, Pierre, Legendre, Louis, 2012.Numerical Ecology. 3rd English ed. Elsevier, Am-sterdam (Chapters 8 and 13).

Levine, Ned, 2008.CrimeStat: a spatial statistical program for the analysis of crime inci-dents. In: Shekhar, Shashi, Xiong, Hui (Eds.), Encyclopedia of GIS. Springer, New York, pp. 187–193.

Loat, Roberto, Petrascheck, Armin, 1997. Consideration of Flood Hazards for Activities with Spatial Impact.Edited by Bundesamt für Wasserwirtschaft (BWW) Federal

Of-fice for Water Management, Bundesamt für Raumplanung (BRP) Federal Office for Spatial Planning, Bundesamt für Umwelt, Wald und Landschaft (BUWAL) Federal

Of-fice for the Environment, Forests and Landscape. Biel. Available online at.http:// www.bafu.admin.ch/publikationen/publikation/00786/index.html?lang=en (checked on 7/21/2016).

Lugeri, Nicola, Kundzewicz, Zbigniew W., Genovese, Elisabetta, Hochrainer, Stefan, Radziejewski, Maciej, 2010. Riverflood risk and adaptation in Europe—assessment of the present status. Mitig. Adapt. Strateg. Glob. Chang. 15 (7):621–639.http://dx. doi.org/10.1007/s11027-009-9211-8.

Lung, Tobias, Lavalle, Carlo, Hiederer, Roland, Dosio, Alessandro, Bouwer, Laurens M., 2013. A multi-hazard regional level impact assessment for Europe combining indica-tors of climatic and non-climatic change. Glob. Environ. Chang. 23 (2):522–536. http://dx.doi.org/10.1016/j.gloenvcha.2012.11.009.

Mazzorana, Bruno, Levaggi, Laura, Keiler, Margreth, Fuchs, Sven, 2012. Towards dynamics inflood risk assessment. Nat. Hazards Earth Syst. Sci. 12 (11):3571–3587.http://dx. doi.org/10.5194/nhess-12-3571-2012.

Merz, Bruno, 2006. Hochwasserrisiken: Grenze und Möglichkeiten der Risikoabschätzung. Schweizerbart'sche Verlagsbuchhandlung, Stuttgart.

Merz, Bruno, Thieken, Annegret H., 2004.Flood risk analysis: concepts and challenges. Österreichische Wasser- und Abfallwirtschaft 56 (3–4), 27–34.

Meyer, Volker, Becker, Nina, Markantonis, Vasileios, Schwarze, Reimund, van den Bergh, Jeroen C.J.M., Bouwer, Laurens M., et al., 2013. Review article: assessing the costs of natural hazards–state of the art and knowledge gaps. Nat. Hazards Earth Syst. Sci. 13 (5):1351–1373.http://dx.doi.org/10.5194/nhess-13-1351-2013.

Mill, John Stuart, 2007.Utilitarianism, Liberty and Representative Government. Wildside Press.

Mori, Koichiro, Perrings, Charles, 2012. Optimal management of theflood risks offl ood-plain development. Sci. Total Environ. 431:109–121.http://dx.doi.org/10.1016/j. scitotenv.2012.04.076.

Muis, Sanne, Guneralp, Burak, Jongman, Brenden, Aerts, Jeroen C.J.H., Ward, Philip J., 2015. Flood risk and adaptation strategies under climate change and urban expansion: a probabilistic analysis using global data. Sci. Total Environ. 538:445–457.http://dx. doi.org/10.1016/j.scitotenv.2015.08.068.

Openshaw, Stan, 1984.The Modifiable Areal Unit Problem. Geo Abstracts Univ. of East An-glia, Norwich (Catmog: concepts and techniques in modern geography). Ord, John K., Getis, Arthur, 1995.Local spatial autocorrelation statistics: distributional

is-sues and an application. Geogr. Anal. Vol. 27 (4), 286–306.

Papathoma-Köhle, Maria, Kappes, Melanie, Keiler, Margreth, Glade, Thomas, 2011. Phys-ical vulnerability assessment for alpine hazards: state of the art and future needs. Nat. Hazards 58 (2):645–680.http://dx.doi.org/10.1007/s11069-010-9632-4.

PLANAT, 2005.Protection Against Natural Hazards in Switzerland - Vision and Strategy (Executive Summary).

Rawls, John, 1971.A theory of justice. Harvard University Press, Cambridge.

Rojas, Rodrigo, Feyen, Luc, Watkiss, Paul, 2013. Climate change and riverfloods in the Eu-ropean Union. Socio-economic consequences and the costs and benefits of adapta-tion. Glob. Environ. Chang. 23 (6):1737–1751. http://dx.doi.org/10.1016/j. gloenvcha.2013.08.006.

Silverman, Bernard W., 1986.Density Estimation for Statistics and Data Analysis. Chap-man and Hall, New York, p. 76 (Monographs on statistics and data analysis). (equa-tion 4.5).

Snow, John, 1855.On the Mode of Communication of Cholera. second ed. Churchill, London.

Staffler, Hanspeter, Pollinger, Rudolf, Zischg, Andreas P., 2008.In: Interpraevent (Ed.), Prioritisation in the planning of permanent protection structures againstfloods and debrisflows on the regional scale in the autonomous province of Bolzano - South Tyrol, pp. 323–334 Conference Proceedings.

Su, Wei, Zhang, Xiaodong, Wang, Zhen, Su, Xiaohui, Huang, Jianxi, Yang, Siquan, Liu, Sanchao, 2011. Analyzing disaster-forming environments and the spatial distribution offlood disasters and snow disasters that occurred in China from 1949 to 2000. Math. Comput. Model. 54 (3–4):1069–1078.http://dx.doi.org/10.1016/j.mcm.2010.11.037. Suter, Hannes, Thomi, Luzius, Kern, Raoul, Künzler, Matthias, Gusterer, Conny, Zischg, Andreas P., et al., 2016.What makes a successfulflood control project? An evaluation of project procedure and risk based on the perspective of Swiss communes. In: Koboltschnig, Gernot (Ed.),Interpraevent 2016.Living With Natural Risks. Proceed-ings of the 13th Congress. Lucerne, pp. 159–167.

swisstopo, 2016a.swissTLM3D, The Topographic Landscape Model TLM (of Switzerland) Version 1.x as per March 2016. (Description and acquisition online at, checked on 9/ 27/2016).

swisstopo, 2016b. Objektkatalog swissTLM3D 1.4.Available online at.https://www. swisstopo.admin.ch/content/swisstopo-internet/de/home/products/landscape/ tlm3d/_jcr_content/contentPar/tabs/items/dokumente/tabPar/downloadlist/ downloadItems/759_1464676126625.download/201603swisstlm3d14okd.pdf (checked on 7/20/2016).

swisstopo, 2016c. swissTLM3D Nachführungsinfo 2016.Available online at.https://www. swisstopo.admin.ch/content/swisstopo-internet/de/home/products/landscape/ tlm3d/_jcr_content/contentPar/tabs/items/dokumente/tabPar/downloadlist_ 159209046/downloadItems/760_1464676201314.download/

swisstlm3drelease2016dedefbarrierefrei.pdf(checked on 7/20/2016).

swisstopo, 2016d. swissBOUNDARIES3D. Grenzen Schweizweit 3D.Available online at. https://www.swisstopo.admin.ch/content/swisstopo-internet/de/home/products/ landscape/boundaries3d/_jcr_content/contentPar/tabs/items/dokumente/tabPar/ downloadlist/downloadItems/793_1464683261852.download/

201602sboundinfodbarrierefrei.pdf(checked on 8/4/2016).

Tango, Toshiro, Takahashi, Kunihiko, 2005. Aflexibly shaped spatial scan statistic for de-tecting clusters. Int. J. Health Geogr. 4:11.http://dx.doi.org/10.1186/1476-072X-4-11. Thaler, Thomas A., Priest, Sally J., Fuchs, Sven, 2016. Evolving inter-regional co-operation inflood risk management. Distances and types of partnership ap-proaches in Austria. Reg. Environ. Chang. 16 (3):841–853.http://dx.doi.org/10. 1007/s10113-015-0796-z.

Thieken, Annegret H., Kienzler, Sarah, Kreibich, Heidi, Kuhlicke, Christian, Kunz, Michael, Mühr, Bernhard, et al., 2016. Review of theflood risk management system in Germa-ny after the majorflood in 2013. E&S 21 (2). http://dx.doi.org/10.5751/ES-08547-210251.

UNISDR, 2015a.Global Assessment Report on Disaster Risk Reduction. Making Develop-ment Sustainable: The Future of Disaster Risk ManageDevelop-ment. Geneva, Switzerland. UNISDR, 2015b.Sendai Framework for Disaster Risk Reduction 2015–2030. Geneva,

Switzerland.

Url, Thomas, Sinabell, Franz, 2008.Flood risk exposure in Austria - options for bearing risk efficiently. Schmollers Jahrbuch. vol. 128, pp. 593–614.

van der Veen, Anne, Logtmeijer, Christian, 2005.Economic hotspots: visualizing vulnera-bility toflooding. Nat. Hazards 36, 65–80.

Visser, Hans, Petersen, Arthur C., Ligtvoet, Willem, 2014. On the relation between weath-er-related disaster impacts, vulnerability and climate change. Clim. Chang. 125 (3–4): 461–477.http://dx.doi.org/10.1007/s10584-014-1179-z.

von Ungern-Sternberg, Thomas, 2004.Efficient Monopolies. The Limits of Competition in the European Property Insurance Market. Oxford University Press, New York. Winsemius, Hessel C., Aerts, Jeroen C.J.H., van Beek, Ludovicus P.H., Bierkens, Marc F.P.,

Bouwman, Arno, Jongman, Brenden, et al., 2016. Global drivers of future riverflood risk. Nat. Clim. Chang. 6 (4):381–385.http://dx.doi.org/10.1038/nclimate2893.

Zimmermann, Markus, Keiler, Margreth, 2015. International frameworks for disaster risk reduction. Useful guidance for sustainable mountain development? Mt. Res. Dev. 35 (2):195–202.http://dx.doi.org/10.1659/MRD-JOURNAL-D-15-00006.1.

Zischg, Andreas P., Schober, Stephan, Sereinig, Norbert, Rauter, Marina, Seymann, Christof, Goldschmidt, Franz, et al., 2013. Monitoring the temporal development of natural hazard risks as a basis indicator for climate change adaptation. Nat. Hazards 67 (3):1045–1058.http://dx.doi.org/10.1007/s11069-011-9927-0.